# Wojciech Zaremba > 1988– · Computer Scientist, Co-founder of OpenAI > > **Recorded contribution:** OpenAI co-founder; robotics and code generation ## How to use this dossier Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory. ## 1. Historical orientation Polish computer scientist Wojciech Zaremba co-founded OpenAI and contributed research on neural program induction, code generation, and robotic learning. His co-authored work includes learning dexterous in-hand manipulation with simulation and domain randomization and the Codex lineage connecting language models to source code. His current role and project portfolio may change; published papers supply the stable evidence. Zaremba's robotics work used domain randomization to make a policy trained in simulation survive the imperfections of a physical robot. Instead of perfectly modeling one hand and object, the simulator varied dynamics, appearance, friction, delays, and other properties so the policy learned behavior robust across a distribution. ## 2. The problem inherited Robots and program-generating models needed to generalize beyond narrow demonstrations despite gaps between simulation and reality and the unforgiving exactness of executable code. ## 3. The central contribution Zaremba co-founded OpenAI and helped lead research applying deep learning to robotics and code, including sim-to-real manipulation and large-scale code models. ## 4. Reconstruct the mechanism 1. Collect simulated trajectories or large corpora of source code paired with context. 2. Train a policy or autoregressive model to predict actions or program tokens. 3. Randomize environments or evaluate generated programs against execution and tests to expose brittle correlations. 4. Transfer to physical hardware or developer workflows with monitoring and human verification. ## 5. What changed downstream - Domain-randomized manipulation advanced modern sim-to-real robot learning. - Code-model research helped establish AI pair-programming as a mass-market application. - The dexterous-hand experiments became a prominent demonstration of sim-to-real transfer, while Codex helped establish source code as a high-value domain for large autoregressive language models. ## 6. Attribution, limits, and uncertainty - The robotics and code results have many co-authors and engineering contributors; Zaremba is not sole creator of Codex or OpenAI robotics. - Generated code can be insecure, incorrect, license-sensitive, or silently unmaintainable; physical transfer can fail outside randomized training support. - Randomization covers only anticipated variation, physical trials still impose safety and hardware costs, and code-generation benchmarks can reward plausible snippets that are insecure, unlicensed, or semantically wrong. ## 7. Reconstruction lab Randomize friction and object mass in a toy simulator, train one control policy, and test held-out combinations. For code, require generated functions to pass adversarial tests and a security review. Remove one randomized variable at a time, identify the physical failure it exposes, and distinguish robustness learned from variation from accidental success on one setup. ## 8. Evidence trail - [Learning Dexterous In-Hand Manipulation](https://arxiv.org/abs/1808.00177) — OpenAI / arXiv - [Evaluating Large Language Models Trained on Code](https://arxiv.org/abs/2107.03374) — OpenAI / arXiv - [Wojciech Zaremba](https://en.wikipedia.org/wiki/Wojciech_Zaremba) — Wikipedia contributors --- *Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*